Selection on Observables

Question 1

Stratum/strat

Answer 1

Groups in which both treated and untreated units share a common characteristic

Question 2

Curse of Dimensionality

Answer 2

As the number of variables increases, it gets harder to find good matches

Question 3

Assignment mechanism

Answer 3

Determines which units receive treatment and which ones don’t. It is also a way to prevent “back-door relationships” between treatment (D) and the outcome (Y)

Question 4

Main estimators

Answer 4

Subclassification, matching, regression, and a combo of the three

Question 5

Propensity score

Answer 5

The selection probability that is conditional on the confound variables. Represented mathematically as p(X) = P(D=1|X). Conditionality on the propensity score is enough to ensure that the treatment indicator and potential outcomes are independent from each other. Can also use this for matching by estimating the propensity with its formula and matching on the estimated propensity score.

Question 6

Fundamental Problem of Causal Inference

Answer 6

Comparisons between the observed outcomes of the treated and untreated groups can be misleading and create bias. So to make casual inferences, you need an explanation for why some units received treatment and why some did not

Question 7

Why use SOO and not an RCT?

Answer 7

SOO is useful for finding causation without clear randomized assignment. RCTs can also be too expensive, unethical or unfeasible to do all the time

Question 8

Identification Assumptions

Answer 8

Treatment Assignment is randomly assigned within each straum of X, and that you can observe participants and nonparticiapnts with the same characteristics.

Question 9

Identifying the ATE

Answer 9

Must know which groups (X) were more likely to receive the treatment, the assignment to treatment has to be random.

Question 10

How to estimate the causal effects of OLS

Answer 10

There are 3 ways: using the subclassification estimator, a regression, or through matching

Question 11

Under what conditions do you need to match?

Answer 11

First, check the balance on pre-treatment covariates, then perform matching, then check the balance in matched data, if the balance is good, then you estimate treatment effects, but if not do the matching again

Question 12

Exact Matching

Answer 12

Match treatment observations to control observations witht he sam characteristics you are observing. The observations need to have the same values to be an exact match.

Question 13

Nearest Neighbor Matching

Answer 13

If perfect matches are not available, then match units with the estimator of the ATT. You can get this estimator by adding the missing potential outcomes of each treat unit using the observed outcomes from the closest untreated unit into the formula for the ATT

Question 14

Subclassification Estimator

Answer 14

Basically dividing the sample into bins based on the value of the key covaraites, then you estimate the difference in means between the treated and untreated units within each of the bins you made, and finally you then construct the weighted averages of the differences in means (by using the proportion of units in each bin to get the ATE, by using the proportion of treated units in each bin to get the ATT, and by using the proportion of untreated units in each bin to recover the ATC).